[Fix] sar_resnet31 TF + PT

[felixdittrich92]

This PR:

• Fixed inference decoding
• Add missing benchmark
• Fixed half precision exception

NOTE:

• with half precision it produces some gibberish looks like an overflow on decoder side (same problem as with vitstr_base) this needs some deeper insight but is not really critical -> other ticket
• I made a small mistake while fixing the SAR architecture:
  I double used the lstmcell on decoder side in pytorch, this explains the small difference in model parameters: TF: 57.2 M --> PT: 55.4 M -> but it does not impact the result quality (as you can see in the benchmarks there is no drop) so i don't see any need to fix and retrain the pt model

Any feedback is welcome 🤗

Closes: #1411 #1441

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 95.82%. Comparing base (709404e) to head (00a99b7).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1513      +/-   ##
==========================================
+ Coverage   95.80%   95.82%   +0.02%     
==========================================
  Files         166      166              
  Lines        7646     7645       -1     
==========================================
+ Hits         7325     7326       +1     
+ Misses        321      319       -2     

Flag	Coverage Δ
unittests	95.82% <100.00%> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.